KML示例

调优思路

通过鲲鹏数学库KML（Kunpeng Math Library）进行优化。

接口描述详见鲲鹏数学库文档。

操作步骤

1. 运行矩阵行列大小为8192的kml_matmult示例。

   ./matmul 8192 6

   返回信息如下：

   Size is 8192, Matrix multiplication method is: 6, Check correctness is: 0
   Initialization time = 2.789213s
   Matrix multiplication time = 0.271790s

   矩阵行列大小为8192，并行计算耗时0.27秒左右。

2. 创建kml_matmult 8192示例的Roofline任务。

   devkit tuner roofline -o kml_matmult_8192 -m region ./matmul 8192 6

   返回信息如下：

   Note:
     1. Roofline task is currently only supported on the 920 platform.
     2. The application must be a binary file in ELF format.
     3. Roofline task collection needs to ensure the application has finished running.
     4. The estimated time of roofline collection is about 3 * application estimated time.
     5. You can learn about the roofline profiling method by looking at document /usr/local/devkit/tuner/docs/ROOFLINE_KNOW_HOW.MD
   RFCOLLECT: Start collection for ./matmul
   RFCOLLECT: Launch application to collect performance metrics of ./matmul
   Size is 8192, Matrix multiplication method is: 6, Check correctness is: 0
   Initialization time = 2.794584s
   ROOFLINE_EVENTS are initialized.
   Matrix multiplication time = 0.432760s
   RFCOLLECT: Launch application to do binary instrumentation of ./matmul
   Size is 8192, Matrix multiplication method is: 6, Check correctness is: 0
   Initialization time = 8.353567s
   Matrix multiplication time = 0.283024s
   RFCOLLECT: Launch benchmarks for measuring roofs
   RFCOLLECT: Processing all collected data
   RFCOLLECT: Result is captured at /matrix_multiplication/rfcollect-20240506-203926.json
   RFCOLLECT: Run "rfreport /matrix_multiplication/rfcollect-20240506-203926.json" to get report.
   
   Get roofline report ...
   The roofline json report: /matrix_multiplication/kml_matmult_8192.json
   The roofline html report: /matrix_multiplication/kml_matmult_8192.html

3. 查看kml_matmult_8192报告。
   图2 kml_matmult_8192报告
   

   获取的roofs的并行度为128，获取到Elapsed Time 0.329s，GFLOP Count 1100.518，Performance 3345.372 GFLOPS。

优化效果

使用KML数学库加速后，计算量恢复到了原有的值（基于数学推导优化，计算量基本不变），数据库优化带来了极大的性能提升，因此端到端性能有了极大的提升，详见下表。

从最基础的并行计算，到最终使用数学库优化， 可以看到程序的端到端时间从516.824秒优化到了0.329秒，整体性能提升1570倍。

• Going UP：Performance提升了3倍多。
• Going RIGHT：较少的右移, 使用向量化指令让计算变得更快，计算密度（FLOP/Byte）没有发生太大的变化 （预期中）。